Methodology

Public data in. The nightly build fetches token and market data from CoinPaprika, daily prices from Binance (and from MEXC, Gate or KuCoin for coins Binance does not list), TVL from DefiLlama, unlock schedules from on-chain vesting contracts (Sablier, Hedgey, UNCX, Team Finance, Streamflow and Jupiter Lock), hacks from DeFiHackLabs' open incident list, contract security from GoPlus and Jupiter, developer activity from GitHub, and listings and delistings from the exchanges' own instrument lists and announcements. It then recomputes every event's price move, every pattern, prediction, setup and score, and writes every page from those numbers. Fresh data goes live around 04:00 UTC.

What the predictions are built from. Price history and a few dated facts: each coin's daily closes and trading volume, Bitcoin's trend, scheduled token unlocks, the coin's size and its type. Today's situation is matched to every similar past moment across every coin we track, and what happened next is counted (how predictions are made).

Checked automatically, every night. The build refuses to publish when its database holds no real data. The full test suite runs on a separate demo database before anything goes live, so live data never decides a test. The export follows every internal link and lists any that break, and a page stays out of search until its own data clears a threshold and most of its text is its own (which pages are indexed).

Dated, and misses kept. Each block names its source and the day its data was fetched or computed. The backtest is re-run every night using only what each past date knew; live predictions are stored the night they are made and scored when their window closes; patterns that failed stay on the Lab's board as failed, and the scorecard prints our hit rate next to the lazy guesses, edge or no edge. Corrections are logged, never edited in silently.

A pattern is a set of conditions on an event and its token at the time: the event kind; the venue; the token's market-cap bucket at the event date (micro under $50M, small $50–500M, mid $500M–$5B, large above $5B, from the nearest daily snapshot); its category group (meme, DEX and DeFi, L1 and L2, other); the unlock's size bucket when the kind is an unlock (under 0.5%, 0.5–2%, over 2% of circulating supply); and Bitcoin's trend (its close above or below its own 50-day average on the event date).

Patterns are built from broad to specific — kind only, kind + venue, kind + size bucket, and so on — and a combination is kept only once at least 8 past cases have a measured next-day move. Shares of cases that were higher carry Wilson 90% intervals, so a thin pattern prints a wide band.

Recent-cases check: a pattern with 20 or more measured cases is re-counted on its older 60% of cases against the newest 40%. When the majority side flips across 50%, the pattern is flagged as flipped on its page and on every chip and card that shows it; older and newer cases disagree and the page says so.

A scheduled event is matched to the most specific stored pattern whose every condition it satisfies; ties go to the larger case count. Pattern pages are indexed from 10 measured cases; thinner patterns are served and linked but marked noindex, like every other thin page on this site.

223 patterns have odds from past cases as of Oct 4, 2026, 125 scheduled or recent events match one, and the pattern index lists every one with all of its past cases.

The same counts at each point (next day, 7, 30 and 90 days, each held to the same 8-case minimum) feed the 0–100 score on every token page, whose arithmetic, weights and backtest are printed on the predictions page. The data behind these pages is published as dated, versioned JSON: /api/predictions.json, /api/scores.json and /api/odds.json, rebuilt with the site.

The situation. On the day of the prediction we describe each coin with a few plain facts, using only data dated on or before that day: is a token unlock of 0.5–2% or more than 2% of the circulating supply scheduled inside the window; is Bitcoin above or below its 50-day average; is the coin up more than 10%, down more than 10% or roughly flat over 30 days, and the same over 7 days at 7%; is it within 20% of its all-time high, 20–60% below, or more than 60% below; did its dollar volume just jump more than two standard deviations above its 30-day normal; its size (under $50M, $50–500M, $500M–$5B, above); and its type (memecoin, DEX and DeFi, layer 1 and 2, other).

The matches. We sample every coin we track every third day across its whole history and keep the past moments that share the same facts, dropping the least important fact until at least 40 matches remain. The chance is the share of those matches that ended higher, pulled toward the overall up-rate of the time so that a small or lucky group cannot print an extreme number. Strong means the chance is at least 10 points from 50%, the case count is 60 or more and its 90% range excludes 50%; lean means at least 5 points; anything closer is a coin flip.

The range. Past moves are compared in units of each coin's own recent volatility and scaled back to this coin's, so a quiet coin gets a narrow cone and a wild one a wide cone. The cone on each token page draws the 10th, 25th, 50th, 75th and 90th percentile of those matched outcomes at the days the engine computes (1, 3, 7, 14, 21 and 30), interpolated in between; the typical move is the 50th percentile, the bad case the 10th and the good case the 90th.

The factors. Under each prediction, "why we think this" lists the facts that moved the chance most: for each one, how often coins in that one situation ended higher against how often coins did overall.

The scorecard. Every week of the past year we re-ran the method using only data available on that date and scored it: how often the direction was right, how often the outcome landed inside the ranges (aim 50% and 80%) and a Brier score against simply using the overall up-rate. Live predictions are stored the night they are made, never edited, and scored separately once their window closes. All of it is on How often are we right?, including where the calls do no better than the base rate.

Setups are separate, named rules re-measured on our own prices every day; each lists its past cases, win rate and whether it held up on its newest cases, at /setups.

Token pages need at least 25 data points, 12 generated sentences and one rich block (three measured events, a security row, an unlock schedule, a DefiLlama slug or a person). A data point is a non-null displayed field or one child row: each event, contract, alias, market, listing, unlock, allocation, raise, hack, audit, repository, person or treasury row counts once.

People need one position on an indexed token and either a description, links or a second position. Exchanges need 10 measured events or 50 markets; investors 3 rounds; auditors 3 audits; creator wallets 3 launches; hubs 5 indexed members; month archives 10 events.

Launch pages are indexed when the token graduated, lived 72 hours, reached a 3x peak, or died with a creator holding three or more launches; other dead launches are served and linked but not indexed, so they never dilute the site.

Each page prints its own score in the footer. The same function feeds the meta robots tag, the sitemaps and the build report.

Outcomes are measured returns, never causes: "the price was 18% lower 30 days after the listing", never "the listing crashed the price". Auditor pages state days between audit and hack and never imply blame.

Wallets carry count-based labels only ("12 launches, 11 dead within 24 hours"). Pulled liquidity is described with the fixed phrase "liquidity was pulled" and never with slang. People are named exactly as the source writes them, with public professional links only.

Facts are compiled from public sources and public blockchain data. Nothing on this site is financial advice.

No correction has been needed yet as of Oct 4, 2026; the first one will be listed here with its date.